CoChat
Sign Up
Cited by 875Ethics and Social Impacts of AIExplainable Artificial Intelligence (XAI)AI and HR Technologies

Fairness and Bias in Artificial Intelligence: A Brief Survey of Sources, Impacts, and Mitigation Strategies

Open access
Ferrara, Emilio·2023·Sci
DOI 10.3390/sci6010003
Open this paper in CoChat — free accountOpen PDF

Abstract

The significant advancements in applying artificial intelligence (AI) to healthcare decision-making, medical diagnosis, and other domains have simultaneously raised concerns about the fairness and bias of AI systems. This is particularly critical in areas like healthcare, employment, criminal justice, credit scoring, and increasingly, in generative AI models (GenAI) that produce synthetic media. Such systems can lead to unfair outcomes and perpetuate existing inequalities, including generative biases that affect the representation of individuals in synthetic data. This survey study offers a succinct, comprehensive overview of fairness and bias in AI, addressing their sources, impacts, and mitigation strategies. We review sources of bias, such as data, algorithm, and human decision biases—highlighting the emergent issue of generative AI bias, where models may reproduce and amplify societal stereotypes. We assess the societal impact of biased AI systems, focusing on perpetuating inequalities and reinforcing harmful stereotypes, especially as generative AI becomes more prevalent in creating content that influences public perception. We explore various proposed mitigation strategies, discuss the ethical considerations of their implementation, and emphasize the need for interdisciplinary collaboration to ensure effectiveness. Through a systematic literature review spanning multiple academic disciplines, we present definitions of AI bias and its different types, including a detailed look at generative AI bias. We discuss the negative impacts of AI bias on individuals and society and provide an overview of current approaches to mitigate AI bias, including data pre-processing, model selection, and post-processing. We emphasize the unique challenges presented by generative AI models and the importance of strategies specifically tailored to address these. Addressing bias in AI requires a holistic approach involving diverse and representative datasets, enhanced transparency and accountability in AI systems, and the exploration of alternative AI paradigms that prioritize fairness and ethical considerations. This survey contributes to the ongoing discussion on developing fair and unbiased AI systems by providing an overview of the sources, impacts, and mitigation strategies related to AI bias, with a particular focus on the emerging field of generative AI.

Verified 2026-10-09 · 1 reference not checked yet

AI snapshot

Written from the paper’s full text.

TL;DR

This survey reviews sources, impacts, and mitigation strategies for fairness and bias in AI, with particular attention to generative AI bias. It argues that bias arises from data, algorithms, and users, can perpetuate inequality, and requires holistic, interdisciplinary mitigation.

Key findings

  • Bias in AI is defined as systematic error in decision-making that leads to unfair outcomes; sources include data bias, algorithmic bias, and user bias, with generative bias as an emergent issue in GenAI outputs.
  • Real-world examples cited include COMPAS being biased against African-American defendants, a healthcare mortality algorithm biased against African-American patients, NIST finding facial recognition less accurate for darker skin tones, and text-to-image models (StableDiffusion, DALL-E, Midjourney) predominantly generating men for CEOs and people of color for criminals or terrorists.
  • Mitigation approaches reviewed include pre-processing data (oversampling, undersampling, synthetic data, adversarial debiasing), model selection (group fairness, individual fairness, regularization, ensemble methods), and post-processing (e.g., equalized odds); GenAI is said to require holistic strategies, audits, monitoring, and feedback loops.
  • Fairness types discussed include group fairness (demographic parity, disparate mistreatment, equal opportunity), individual fairness, counterfactual fairness, procedural fairness, and causal fairness; the paper distinguishes fairness as a deliberate goal from bias, which can be unintentional.

Methods

  • Design: a survey described as a systematic literature review spanning multiple academic disciplines, providing an overview of AI bias sources, impacts, and mitigation strategies.
  • Data or sample: no primary dataset or sample is analyzed; the paper draws on prior literature and illustrative examples from healthcare, criminal justice, hiring, credit scoring, facial recognition, and generative AI.
  • Main analysis: narrative synthesis, with tables characterizing types of AI bias, mitigation approaches and challenges, and definitions of fairness; no quantitative meta-analysis or formal search protocol is reported.

Limitations

  • Stated: lack of diverse and representative training data; difficulty identifying and measuring different types of bias; trade-offs between fairness and accuracy; ethical questions about prioritizing types of bias or groups; and limitations of mitigation methods, including pre-processing being time-consuming or ineffective if data are already biased, model selection lacking consensus on fairness, and post-processing being complex and data-intensive.
  • Stated: fairness definitions vary and may conflict; group and individual fairness approaches have trade-offs; fairness metrics may not capture intersectionality; and mitigation attempts can have unintended consequences, such as predictive policing efforts potentially increasing racial disparities in arrests.
  • Evident: as a brief survey, the paper does not report a systematic search strategy, inclusion/exclusion criteria, or quantitative synthesis; its examples are illustrative rather than exhaustive.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/sci6010003

Cites

48 of 50 references verified0 citation errors · 1 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 65 cited sources checked · 0 supported · 0 flagged

  • Findingcitation

    Model selection methods may be limited by the lack of consensus on what constitutes fairness.

    “Additionally, model selection methods may be limited by the lack of consensus on what constitutes fairness, and post-processing methods can be complex and require large amounts of additional data [32].”
    Quote in the paperNot a substantial claim
  • Findingreasoning

    The lack of diverse and representative training data is a main challenge to mitigating bias in AI.

    “One of the main challenges is the lack of diverse and representative training data.”
    Quote in the paperSubstantial claim
  • Premisecitation

    Biased algorithms can lead to the underrepresentation of certain groups, such as people of color or those from lower socioeconomic backgrounds, in credit scoring systems, making it harder for them to access loans or mortgages.

    “Biased algorithms can lead to the underrepresentation of certain groups, such as people of color or those from lower socioeconomic backgrounds, in credit scoring systems, making it harder for them to access loans or mortgages [25].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25] Dwork, C., Hardt, M., Pitassi, T., Reingold, O., and Zemel, R. (2012, January 8–10). Fairness through awareness. Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, Cambridge, MA, USA. source

      Not checked in the free run.

  • Premisecitation

    The same GenAI models that exhibit gender bias also portray criminals or terrorists as people of color.

    “The same GenAI models that exhibit gender bias, perhaps unsurprisingly, also portray criminals or terrorists as people of color.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23] Mittelstadt, 2016, The ethics of algorithms: Mapping the debate, Big Data Soc. source

      Not checked in the free run.

  • Interpretationreasoning

    GenAI models trained on internet-sourced images likely suffer from bias because the data mirrored existing disparities.

    “GenAI models trained on internet-sourced images likely suffer from this bias, as the data mirrored existing disparities.”
    Quote in the paperSubstantial claim
  • May combine two claims. 26
  • Classified as probably not a substantial claim. 24
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingnone given

    Pre-processing data can be time-consuming and may not always be effective, especially if the data used to train models are already biased.

    “For example, pre-processing data can be time-consuming and may not always be effective, especially if the data used to train models are already biased.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Algorithmic bias can be difficult to detect and quantify, especially when the algorithms are complex or opaque.

    “Algorithmic bias can be difficult to detect and quantify, especially when the algorithms are complex or opaque.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    The sources of bias may be difficult to isolate, as bias can arise from multiple sources, such as the data, the algorithm, and the user.

    “Additionally, the sources of bias may be difficult to isolate, as bias can arise from multiple sources, such as the data, the algorithm, and the user.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Collecting diverse and representative data can be challenging, especially when dealing with sensitive or rare events.

    “However, collecting diverse and representative data can be challenging, especially when dealing with sensitive or rare events.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingreasoning

    Privacy concerns when collecting certain types of data, such as medical records or financial information, can limit the effectiveness of dataset augmentation as a mitigation approach.

    “Additionally, there may be privacy concerns when collecting certain types of data, such as medical records or financial information.”
    • Also stated in: § 2. Unintended consequences on the distribution of outcomes · plain
      “These challenges can limit the effectiveness of dataset augmentation as a mitigation approach.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    These biases can perpetuate systemic discrimination and inequality, with detrimental effects on individuals and communities in areas like hiring, lending, and criminal justice.

    “These biases can perpetuate systemic discrimination and inequality, with detrimental effects on individuals and communities in areas like hiring, lending, and criminal justice [3][4][5][6][7].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3][4][5][6][7]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationcitation

    In generative AI, addressing bias requires a holistic strategy.

    “In the realm of generative AI, addressing bias is even more challenging as it requires a holistic strategy [14,21,22].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [14] Ferrara, E. (2023). Should ChatGPT be Biased? Challenges and Risks of Bias in Large Language Models. First Monday, 28. source

      Not checked in the free run.

    • cites [21] Ferrara, E. (2023). GenAI against humanity: Nefarious applications of generative artificial intelligence and large language models. arXiv. source

      Not checked in the free run.

    • cites [22] Ferrara, E. (2023). The butterfly effect in artificial intelligence systems: Implications for AI bias and fairness. arXiv. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    An AI system used in a hiring process may disproportionately exclude candidates from marginalized groups, limiting their ability to access employment opportunities.

    “For example, an AI system used in a hiring process may disproportionately exclude candidates from marginalized groups, limiting their ability to access employment opportunities and contribute to society.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    AI systems can lead to unfair outcomes and perpetuate existing inequalities, including generative biases that affect the representation of individuals in synthetic data.

    “Such systems can lead to unfair outcomes and perpetuate existing inequalities, including generative biases that affect the representation of individuals in synthetic data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The importance of fairness and bias in AI is widely recognized by researchers, policymakers, and the academic community.

    “The importance of fairness and bias in AI is widely recognized by researchers, policymakers, and the academic community [1,[12][13][14][15][16].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1,[12][13][14][15][16]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    Bias in AI can perpetuate and even amplify existing inequalities, leading to discrimination against marginalized groups and limiting their access to essential services.

    “Bias in AI can perpetuate and even amplify existing inequalities, leading to discrimination against marginalized groups and limiting their access to essential services.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Facial recognition algorithms trained on data primarily consisting of men can struggle to recognize female faces accurately, perpetuating gender bias in security systems.

    “For instance, facial recognition algorithms trained on data primarily consisting of men can struggle to recognize female faces accurately, perpetuating gender bias in security systems [1].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1] Buolamwini, J., and Gebru, T. (2018, January 23–24). Gender shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of the 1st Conference on Fairness, Accountability and Transparency, New York, NY, USA.

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    Biased AI systems can limit individual freedoms and reinforce societal power dynamics.

    “When AI systems are biased, they can limit individual freedoms and reinforce societal power dynamics.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Bias in AI can lead to new forms of discrimination based on skin color, ethnicity, or physical appearance.

    “In addition to perpetuating gender stereotypes and discrimination, it can also lead to new forms of discrimination based on skin color, ethnicity, or physical appearance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Post-processing methods can be complex and require large amounts of additional data.

    “Additionally, model selection methods may be limited by the lack of consensus on what constitutes fairness, and post-processing methods can be complex and require large amounts of additional data [32].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] Barocas, 2016, Big data’s disparate impact, Calif. Law Rev.

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    The widespread use of biased AI systems can entrench discriminatory narratives and hinder efforts toward equality and inclusivity.

    “On a societal level, the widespread use of such biased AI systems can entrench discriminatory narratives and hinder efforts toward equality and inclusivity.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Machine learning models can learn and replicate patterns of bias present in the data used to train them, resulting in unfair or discriminatory outcomes.

    “Machine learning models, a type of AI system, can learn and replicate patterns of bias present in the data used to train them, resulting in unfair or discriminatory outcomes.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    The use of biased AI systems may undermine public trust in technology, leading to decreased adoption and even rejection of new technologies.

    “Moreover, the use of biased AI systems may undermine public trust in technology, leading to decreased adoption and even rejection of new technologies.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Bias in AI can arise from various sources, including data collection, algorithm design, and human interpretation.

    “In the context of AI, bias can arise from various sources, including data collection, algorithm design, and human interpretation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The COMPAS system used in the United States criminal justice system was biased against African-American defendants, who were more likely to be labeled as high-risk even if they had no prior convictions.

    “A study by ProPublica found that the system was biased against African-American defendants, as they were more likely to be labeled as high-risk even if they had no prior convictions.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites ProPublica

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Biased algorithms used in the criminal justice system can lead to unfair treatment of certain groups, particularly people of color, who are more likely to be wrongly convicted or receive harsher sentences.

    “For example, biased algorithms used in the criminal justice system can lead to unfair treatment of certain groups, particularly people of color, who are more likely to be wrongly convicted or receive harsher sentences [1].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [1] Buolamwini, J., and Gebru, T. (2018, January 23–24). Gender shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of the 1st Conference on Fairness, Accountability and Transparency, New York, NY, USA.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Different people and groups may have different definitions of fairness, and these definitions may change over time.

    “Different people and groups may have different definitions of fairness, and these definitions may change over time [39].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [39] Ezzeldin, Y.H., Yan, S., He, C., Ferrara, E., and Avestimehr, S. (2023, January 7–14). Fairfed: Enabling group fairness in federated learning. Proceedings of the AAAI 2023—37th AAAI Conference on Artificial Intelligence, Washington, DC, USA.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Regularization penalizes models for making discriminatory predictions, and ensemble methods combine multiple models to reduce bias.

    “This can be performed through techniques such as regularization, which penalizes models for making discriminatory predictions, or through ensemble methods, which combine multiple models to reduce bias [25].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [25] Dwork, C., Hardt, M., Pitassi, T., Reingold, O., and Zemel, R. (2012, January 8–10). Fairness through awareness. Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, Cambridge, MA, USA. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Other approaches to ensuring fairness include transparency, accountability, and explainability.

    “Other approaches include transparency, accountability, and explainability.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [26] Ananny, 2018, Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability, New Media Soc. source

      Not checked in the free run.

    • cites [38] Donovan, J., Caplan, R., Matthews, J., and Hanson, L. (2018). Algorithmic Accountability: A Primer, Data & Society.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Achieving both fairness and accuracy can be challenging and requires careful consideration of the trade-offs involved.

    “Achieving both fairness and accuracy can be challenging and requires careful consideration of the trade-offs involved.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Different types of fairness may conflict with each other, and trade-offs may need to be made to achieve fairness in specific contexts.

    “Additionally, different types of fairness may conflict with each other, and trade-offs may need to be made to achieve fairness in specific contexts [26].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26] Ananny, 2018, Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability, New Media Soc. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    There are several types of fairness proposed in the literature, including group fairness, individual fairness, and counterfactual fairness.

    “There are several types of fairness proposed in the literature, including group fairness, individual fairness, and counterfactual fairness.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    It may be difficult to determine which types of fairness are most appropriate for a given context and how to balance them appropriately.

    “Additionally, it may be difficult to determine which types of fairness are most appropriate for a given context and how to balance them appropriately [4][5][6].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [4] Kleinberg, 2018, Human decisions and machine predictions, Q. J. Econ.

      Not checked in the free run.

    • cites [5] Kleinberg, 2018, Discrimination in the Age of Algorithms, J. Leg. Anal. source

      Not checked in the free run.

    • cites [6] Kleinberg, 2018, Algorithmic fairness, AEA Pap. Proc. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The Northpointe COMPAS was modified to include a race-neutral version of the algorithm that achieved similar accuracy while reducing racial bias.

    “To address this issue, the Northpointe COMPAS was modified to include a “race-neutral” version of the algorithm that achieved similar accuracy while reducing racial bias [19].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Angwin, J., Larson, J., Mattu, S., and Kirchner, L. (2016). Ethics of Data and Analytics, Auerbach Publications.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Researchers have proposed using techniques such as subgroup analysis to identify and address bias in the data used to train AI models.

    “To address this issue, researchers have proposed using techniques such as subgroup analysis to identify and address bias in the data used to train AI models [11].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [11] Yan, S., Kao, H.T., and Ferrara, E. (July, January 26). Fair class balancing: Enhancing model fairness without observing sensitive attributes. Proceedings of the 29th ACM International Conference on Information & Knowledge Management, Online.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Bias can be viewed as a technical issue, while fairness is a social and ethical issue.

    “In other words, bias can be viewed as a technical issue, while fairness is a social and ethical issue [26].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26] Ananny, 2018, Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability, New Media Soc. source

      Not checked in the free run.

    • Rests on an assumption: Bias can be unintentional, while fairness is inherently a deliberate and intentional goal.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationnone given

    Different types of fairness are not mutually exclusive and may overlap in practice.

    “It is important to note that these different types of fairness are not mutually exclusive and may overlap in practice.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Several different types of fairness have been proposed in the literature, including group fairness, individual fairness, and counterfactual fairness.

    “There are several different types of fairness that have been proposed in the literature, including group fairness, individual fairness, and counterfactual fairness [30].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30] Zafar, M.B., Valera, I., Gomez Rodriguez, M., and Gummadi, K.P. (2017, January 3–7). Fairness beyond disparate treatment & disparate impact: Learning classification without disparate mistreatment. Proceedings of the 26th International Conference on World Wide Web, Perth, Australia. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The COMPAS system is used to predict the likelihood of recidivism in criminal defendants.

    “One example is the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) system, which is used to predict the likelihood of recidivism in criminal defendants.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Angwin, J., Larson, J., Mattu, S., and Kirchner, L. (2016). Ethics of Data and Analytics, Auerbach Publications.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationcitation

    Diversifying the teams involved in AI development and evaluation is crucial, as it brings a multiplicity of perspectives that can better identify and correct for biases.

    “Diversifying the teams involved in AI development and evaluation is equally crucial, as it brings a multiplicity of perspectives that can better identify and correct for biases [55][56][57].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [55] Chauhan, 2022, The Role of Data and Artificial Intelligence in Driving Diversity, Equity, and Inclusion, Computer source

      Not checked in the free run.

    • cites [56] Holstein, K., Wortman Vaughan, J., Daumé, H., Dudik, M., and Wallach, H. (2019, January 4–9). Improving fairness in machine learning systems: What do industry practitioners need?. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, Glasgow, UK. source

      Not checked in the free run.

    • cites [57] Stathoulopoulos, K., Mateos-Garcia, J.C., and Owen, H. (2023, December 15). Gender Diversity in AI Research. Available online: https://www.nesta.org.uk/report/gender-diversity-ai/.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Bias can be either positive or negative, whereas fairness is only concerned with negative bias or discrimination.

    “Another difference is that bias can be either positive or negative, whereas fairness is only concerned with negative bias or discrimination [36].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36] 2017, Measuring discrimination in algorithmic decision making, Data Min. Knowl. Discov. source

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationcitation

    AI raises ethical implications for preserving privacy and requires transparency, oversight, and continuous evaluation.

    “We also considered the ethical implications of AI in preserving privacy and the necessity for transparency, oversight, and continuous evaluation of AI systems [52][53][54].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [52] Lipton, 2018, The mythos of model interpretability: In machine learning, the concept of interpretability is both important and slippery, Queue source

      Not checked in the free run.

    • cites [53] Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I.D., and Gebru, T. (2019, January 29–31). Model cards for model reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency, Atlanta, GA, USA. source

      Not checked in the free run.

    • cites [54] Raji, I.D., and Buolamwini, J. (2019, January 29–31). Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial AI products. Proceedings of the Conference on Fairness, Accountability, and Transparency, Atlanta, GA, USA. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The COMPAS system was biased against African-American defendants, as it was more likely to falsely predict that they would re-offend compared to white defendants.

    “Research has shown that the system was biased against African-American defendants, as it was more likely to falsely predict that they would re-offend compared to white defendants [19].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Angwin, J., Larson, J., Mattu, S., and Kirchner, L. (2016). Ethics of Data and Analytics, Auerbach Publications.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Ethical considerations around how to prioritize different types of bias and which groups to prioritize add complexity to the development and implementation of bias mitigation approaches.

    “These ethical considerations can add complexity to the development and implementation of bias mitigation approaches.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Bias is not the only factor that can lead to unfairness, and achieving fairness may require additional efforts beyond bias mitigation.

    “However, it is important to recognize that bias is not the only factor that can lead to unfairness, and achieving fairness may require additional efforts beyond bias mitigation [27].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [27] Corbett-Davies, S., Pierson, E., Feller, A., Goel, S., and Huq, A. (2017, January 13–17). Algorithmic decision making and the cost of fairness. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada. source

      Not checked in the free run.

    • Rests on an assumption: Addressing bias in training data or algorithms can help reduce the likelihood of unfair outcomes.

  • May combine two claims.

    Interpretationreasoning

    Current approaches to ensuring fairness in AI have shown promising results but are not without limitations and challenges.

    “While these approaches have shown promising results in promoting fairness in AI, they are not without limitations and challenges.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Many current approaches to ensuring fairness in AI rely on statistical methods and assumptions that may not accurately capture the complexity of human behavior and decision-making.
41 of 41

  • The paper relies on this without support.

    The use of AI systems is growing.

    “The growing use of AI systems has intensified discussions around fairness and bias in artificial intelligence as potential biases and discrimination become more apparent.”

  • The paper relies on this without support.

    Generative AI is becoming more prevalent in creating content that influences public perception.

    “especially as generative AI becomes more prevalent in creating content that influences public perception.”

  • The paper relies on this without support.

    Bias can be unintentional, while fairness is inherently a deliberate and intentional goal.

    “One key difference between fairness and bias is that while bias can be unintentional, fairness is inherently a deliberate and intentional goal.”

  • The paper relies on this without support.

    Addressing bias in training data or algorithms can help reduce the likelihood of unfair outcomes.

    “For example, addressing bias in training data or algorithms can help reduce the likelihood of unfair outcomes.”

  • The paper relies on this without support.

    Group fairness can be achieved through various techniques such as re-sampling, pre-processing, or post-processing of the data used to train the AI model.

    “This can be achieved through various techniques such as re-sampling, pre-processing, or postprocessing of the data used to train the AI model.”

  • The paper relies on this without support.

    Individual fairness can be achieved through techniques such as counterfactual fairness or causal fairness.

    “Individual fairness can be achieved through techniques such as counterfactual fairness or causal fairness.”

  • The paper relies on this without support.

    Many current approaches to ensuring fairness in AI rely on statistical methods and assumptions that may not accurately capture the complexity of human behavior and decision-making.

    “Furthermore, many of the current approaches to ensuring fairness in AI rely on statistical methods and assumptions that may not accurately capture the complexity of human behavior and decision-making.”

  • The paper relies on this without support.

    Bias can arise due to various factors, such as biased data or algorithmic design.

    “Bias can arise due to various factors, such as biased data or algorithmic design, but fairness requires a conscious effort to ensure that the algorithm does not discriminate against any group or individual.”

  • The paper relies on this without support.

    Fairness requires a conscious effort to ensure that the algorithm does not discriminate against any group or individual.

    “Bias can arise due to various factors, such as biased data or algorithmic design, but fairness requires a conscious effort to ensure that the algorithm does not discriminate against any group or individual.”

  • The paper relies on this without support.

    If an AI model is trained on data that are biased toward a particular group, re-sampling techniques can be used to create a balanced dataset where each group is represented equally.

    “For example, if an AI model is trained on data that are biased toward a particular group, re-sampling techniques can be used to create a balanced dataset where each group is represented equally.”

  • The paper relies on this without support.

    Pre-processing or post-processing can be used to adjust the output of the AI model to ensure that it does not unfairly disadvantage any group.

    “Other techniques, such as pre-processing or post-processing, can be used to adjust the output of the AI model to ensure that it does not unfairly disadvantage any group.”
11 of 11

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Fairness both excludes and includes favoritism

    Minor

    The paper says fairness is concerned only with negative bias, but elsewhere defines fairness as the absence of discrimination or favoritism. It also defines positive bias as systematically favoring a group, so the stated scope of fairness conflicts with its definition.

  • Model training classified as pre-processing

    Minor

    The paper defines pre-processing as addressing bias before the model is trained, then lists a technique involving training the model as an example of pre-processing.

2 of 2

Cites

Flagged first, then by citations

9 of the 59 references couldn’t be matched to a known paper.

Research works printed but not linked: 3

  1. Eubanks, V. (2018). Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor, St. Martin’s Press.
  2. Corbett-Davies, S., and Goel, S. (2018). The measure and mismeasure of fairness: A critical review of fair machine learning. arXiv.
  3. Crenshaw, K. (1989). Feminist Legal Theories, Routledge.

Cited by

Sorted by citations

Related

Authors
Ferrara, E.
Affiliations
University of Southern California (US)
Corresponding author
Ferrara, Emilio
ORCID
Ferrara, Emilio
Venue
Sci
Publisher
MDPI AG
Published
2023-12-26
Volume
6
Issue
1
Pages
3
ISSN
2413-4155 (ISSN-L)
DOI
10.3390/sci6010003
OpenAlex
W4390229867
Semantic Scholar
53b04ccd2a001467d7ce168e9ce20b16a9466a69
Cited by
875
References
59
Citations by year
2026: 271 · 2025: 433 · 2024: 162 · 2023: 6 · 2022: 2 · 2020: 1
FWCI
218.36
Type
journal article
Language
en
Open access
Licence
Locations
  • Sci (publishedVersion, cc-by)
  • SSRN Electronic Journal (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University) (cc-by)
Topics
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Business, Management and Accounting › Organizational Behavior and Human Resource Management › AI and HR Technologies
Keywords
algorithmic fairness, algorithmic bias, generative AI bias, generative AI, synthetic media, data bias, bias mitigation, bias mitigation strategies, data preprocessing, model selection, post-processing, societal impact of AI, stereotypes, artificial intelligence in healthcare, AI ethics

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review